Justifiability and AI: putting explainability in its place

Area of Science:

• Computer Science
• Philosophy of Technology
• Artificial Intelligence Ethics

Background:

• The increasing integration of artificial intelligence and machine learning (AI/ML) systems into critical societal domains like healthcare and finance highlights the significant challenge posed by their inherent opacity.
• Existing post hoc explainability algorithms, designed to demystify AI/ML decision-making, face criticism for offering mere rationalizations rather than genuine insights into the model's reasoning.

Purpose of the Study:

• To introduce and defend the concept of AI/ML justifiability as a more robust framework for evaluating AI/ML systems than post hoc explainability.
• To explore the philosophical underpinnings of justifiability, distinguishing between motivating and normative reasons in the context of AI/ML decision-making.

Main Methods:

• Conceptual analysis drawing from the philosophy of action to differentiate between motivating and normative reasons.
• Argumentation for the superiority of justifiability over explainability in contexts demanding trust and accountability.

Main Results:

• Effective AI/ML justifications can be achieved using normative reasons alone, which are within the capacity of AI/ML systems to provide.
• Post hoc explanations, while attempting to provide motivating reasons, are often insufficient and can be misleading.

Conclusions:

• AI/ML systems can be deemed trustworthy based on their justifiability, irrespective of whether their internal workings are fully explainable.
• The pursuit of AI/ML justifiability offers a promising direction for developing reliable and accountable AI/ML applications in sensitive areas.